The present invention relates to a switch-connected HyperX network system, and more particularly to, a large-scale switch-connected network system including multiple HyperX network planes and a method for building the large-scale switch-connected network system.
Two-dimensional (2D) all-to-all connected network such as HyperX network topologies have a low diameter and a good all-to-all communication bandwidth. It exploits all-to-all wiring to achieve these benefits with low cost. For example, such 2D HyperX network topologies can be used by a large single multiprocessor (SMPs) such as Power8® (registered trademark of International Business Machines Corporation) 16 socket SMPs.
There are several limiting issues in deploying a network system with the 2D HyperX network topology. In a high-performance computing (HPC) system, compute resources are typically running partitioned. The all-to-all connected network system can be divided into multiple partitions used for different jobs. For example, when the all-to-all connected network system is divided into 2 equally-sized partitions for different independent jobs, half of the original all-to-all links may become inter-partition idle links. As a result, half of the network bandwidth will be lost. These inter-partition links may be used for intra-partition communication by means of indirect routing, however this may cause undesirable inter-job interference.
In addition, with the all-to-all connected network system, there may be limitation in scaling up a network size by adding new nodes and switches to the existing all-to-all network system. To add a new component (e.g., node, switch, and network plane), the existing components may be rewired to maintain the all-to-all wiring.